feat(ascend): integrate InfiniOps backend

[baominghelly]

Summary

• register Ascend devices for InfiniOps-backed operator dispatch
• enable Ascend RoPE cache initialization and rotary embedding dispatch
• select the native Ascend Argmax implementation
• rely on the existing INFINI_OPS_OPS adapter whitelist instead of an Ascend-specific remove_files list

Testing

Ascend 910C, CANN 9.0.0, InfiniLM main, Qwen3-0.6B:

• clean xmake configuration with installed InfiniOps and InfiniRT packages
• xmake build -r -j16 infinicore_cpp_api
• xmake build -r -j16 _infinicore
• whitelist build-log check:
  • filtered: conv2d, gelu, gelutanh, paged_caching, sigmoid, softmax, topksoftmax, random_sample
  • selected: rotary_embedding, causal_softmax
• Python import validation with torch_npu: NPU available, one visible device
• InfiniLM _infinilm clean rebuild against this InfiniCore tree
• Qwen3-0.6B single-device inference passed with INFINI_OPS_TRACE_CALLS=1
  • generated text successfully
  • 6,768 InfiniOps calls observed, including Embedding, RmsNorm, Gemm, RotaryEmbedding, CausalSoftmax, and Add
• git diff --check

Dependencies

Requires an installed InfiniOps package with the corresponding Ascend operator support.